CN102880866A - Method for extracting face features - Google Patents

Abstract

The invention discloses a method for extracting face feature. Accordingly, body posture analytical data and depth data provided by a Kinect camera are combined with a Depth-Active Appearance Model (AMM) algorithm, and the method based on 2.5 dimensional images is formed. The method comprises steps of training the AMM of the Depth-AMM algorithm by using a principal component analysis method and extracting face features based on the AMM of the Depth-AMM algorithm after completed training.

Description

A kind of face feature extraction method

Technical field

The present invention relates to the image analysis technology field, specifically is a kind of face feature extraction method.

Background technology

Facial Feature Extraction Technology is exactly the accurate location of automatically orienting each organ of people's face by computing machine on a width of cloth facial image, and all need the position of extract minutiae comprising eyes, nose, face and people's face outline etc.Face characteristic extracts can provide corresponding basic data for research work such as recognition of face, expression posture analysis, face trackings.Current many feature extraction algorithms such as main signature analysis (PCA), local binary (LBP), linear discriminant analysis (LDA), the Gabor wavelet transformations etc. of existing can be used for extracting face characteristic, can be than good berth but these methods can only be descended at specified conditions (light, posture, cosmetic and countenance are appropriate), and what obtain all is information some bottoms, complicated, is used for the effect that recognition of face and cluster are difficult to obtain.

Initiatively by being applied in a lot of fields of success, there is the modeling of people's face in the field that relates to apparent model (Active Appearance Model, AAM), the human eye modeling, Facial expression recognition, image segmentation and analysis, attitude is estimated, face tracking and gesture identification etc.The face characteristic location algorithm roughly can be divided into two classes according to the usage data dimension: the face characteristic based on two dimensional image is located, is located based on the face characteristic of 3-D view.Locate because existing people's face detects the inherent limitations of cutting techniques based on the face characteristic of two dimensional image, illumination, background and personage's attitude etc. have larger impact to result.Used expensive triplex scanner to be used for the generating three-dimensional facial image based on the face characteristic of 3-D view location, used curvature to calculate and the algorithm of global registration, to computing machine and treatment facility require too highly, be difficult to practical popularization.Initiatively apparent model (AAM) is exactly to use a kind of method of locating based on the face characteristic of two dimensional image comparatively widely.

Kinect camera 2010 goes on the market in the U.S., this small and exquisite cheap depth camera equipment, also can reach on the commercial hardware more than per second 200 frames, under complex background and personage's attitude condition, can accurately follow the tracks of and cut apart face image, bring a series of revolutionary variations for the fields such as computer vision, computer graphics, man-machine interaction.As can be known aforementioned based on this section, although the Kinect camera has been widely used in human body attitude analysis identification, and under complex background and personage's attitude condition, can accurately follow the tracks of and cut apart face image, do not come method that face characteristic is positioned but also utilize so far human body attitude that the Kinect camera provides to analyze data and depth data, yet there are no namely that the human body attitude that utilizes the Kinect camera to provide is analyzed data and depth data carries out the method that face characteristic extracts.

Summary of the invention

Technical matters to be solved by this invention is, provides a kind of human body attitude that the Kinect camera is provided to analyze data and depth data is fused in the Depth-AAM algorithm, forms the face feature extraction method based on 2.5 dimension images.

Technical scheme of the present invention is, a kind of face feature extraction method is provided, may further comprise the steps,

1) adopt the principal component analysis method to train the apparent model of Depth-AAM algorithm:

1. utilize the training of Kinect collected by camera texture image and the depth image of facial image, depth image is compressed to 0 ~ 255 pixel coverage from 0 ~ 65535 pixel coverage, the α passage of substitution four-way image, merge into RGBD four-way image with texture image again, and it is carried out craft demarcate several point;

2. define people's face shape for forming v apex coordinate s=(x of grid ₁, y ₁..., x _v, y _v) ^TThe shape vector that the summit consists of is set up the two-dimensional linear model with the principal component analysis method, and shape vector is expressed as basic configuration s ₀Add m shape vector s _iLinear combination

P=(p ₁..., p _m) ^TThe feature value vector of form matrix, s ₀Be the standard attitude of facial image, s _iEigenwert p _iThe characteristic of correspondence vector;

3. with s ₀, RGBD four-way image I _iWith its corresponding manual markings s _i ^*Be transformed into RGBD four-way facial image under the standard attitude, i.e. s with the piecemeal affined transformation ₀With s _i' triangle gridding corresponding one by one, piecemeal affined transformation expression formula is x'=a ₁X+a ₂Y+a ₃And y'=b ₁X+b ₂Y+b ₃, (x, y) is s ₀A upper coordinate, (x ', y ') be s _i' upper the coordinate corresponding with (x, y), a ₁With b ₂Be the convergent-divergent yardstick of directions X and Y-direction, a ₂And b ₁Be rotation yardstick, a ₃And b ₃Translation size for directions X and Y-direction adopts the method for undetermined coefficients to obtain corresponding parameter (a ₁, a ₂, a ₃, b ₁, b ₂, b ₃);

4. 3. all training facial images are gone on foot conversion through the, obtain its facial image I under the standard attitude _i', and adopt the principal component analysis method

Here λ _iThe parameter of i apparent vector, apparent parameter vector λ={ λ ₁, λ ₂..., λ _nThat input picture is corresponding to the eigenwert of the apparent parameter of this AAM model, to represent the full detail of input picture, i apparent vectorial A _i(x) corresponding to the large eigenwert of i in the apparent parameter vector;

By 1. 2. 3. 4. the step trains the eigenwert that obtains each appearance features namely to finish the training of the apparent model of Depth-AAM algorithm;

The apparent model of the Depth-AAM algorithm of 2) finishing based on training carries out face characteristic and extracts:

5. the Kinect camera adopts the API of Kinect winsdk to be partitioned into human body image, and obtains head node location coordinate, head joint rotation direction θ and degree of confidence Conf thereof according to the human body depth image _θAccording to the human body depth image, human region is put white, zone outside the human body is put black, node location constantly enlarges the hunting zone about up and down simultaneously from the head, when all reaching black border about upper, stops the search of lower boundary, and the maximum region of definite people's face, its top left corner apex coordinate is designated as (x _HeadLU, y _HeadLU), zone length and width are designated as (length _Head, width _Head);

6. the global shape parameter q is defined as $N(x;q)=\left[\begin{array}{cc}1+a& -b\\ b& 1+a\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]+\left[\begin{array}{c}{t}_x\\ t_y\end{array}\right],$ Parameter (a, b) is expressed as a=kcos θ-and b=ksin θ, (t _x, t _y) be the translation of directions X and Y-direction, in order to write conveniently (a, b, t _x, t _y) be designated as (q ₁, q ₂, q ₃, q ₄) be the global shape parameter q;

7. the objective function of Depth-AAM algorithm match is the absolute value of input picture and formal synthesis image difference $E=\underset{x\∈s_0}{\mathrm{\Σ}}[A_0\left(x\right)+\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}_i{A}_i\left(x\right)-I\left(N(W(x;p);q)\right)],$ Rotation direction θ initialization q is used in primary Depth-AAM match ₁, q ₂, the people's face positional information initialization q that obtains in the use 5. ₃, q ₄P represents the standard attitude after the initialization, finding the solution parameter p and q makes the coordinate figure of the poor described point that hour obtains of image energy and apparent parameter vector namely finish face characteristic to extract, be specially, to parameter p and q differentiate, try to achieve variation delta p and the Δ q of parameter p and q, it is poor that iteration is asked for minimum image energy $\mathrm{\Δq}=H^{-1}\underset{i=0}{\overset{n}{\mathrm{\Σ}}}S{D}_i[I\left(N(W(x;p);q)\right)-A_0\left(x\right)],$ The scope of i is i=1 ..., 4, $\mathrm{\Δp}=H^{-1}\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{\mathrm{SD}}_{j+4}[I\left(N(W(x;p);q)\right)-A_0\left(x\right)],$ The scope of j is j=1. .., and, wherein

The scope of k is k=1 ..., 4,

The scope of l is l=1 ..., 68, H is the Hessian matrix, $H=\underset{x\∈S_0}{\mathrm{\Σ}}{\left[{\mathrm{SD}}_k\left(x\right)\right]}^T\left[{\mathrm{SD}}_k\left(x\right)\right].$

Adopt the pyramid algorith of layering, will cut apart the facial image that obtains and narrow down to

With

Right first

The target facial image of size carries out primary Depth-AAM match, obtains rough form parameter p ¹With overall deformation parameter q, with primary Depth-AAM fitted shapes parameter p ¹Amplify 2 times, substitution Depth-AAM carries out the match second time and obtains form parameter p ²With overall deformation parameter q ², again with p ²Amplify 2 times, substitution Depth-AAM carries out for the third time match and obtains p ³, p ³68 point coordinates that namely match obtains, vectorial λ is exactly apparent parameter; Described pyramid algorith is divided into 3 layers, and every layer of pyramid maximum iteration time is 30 times, if the modulus of the difference of 2 p in front and back is considered as iteration convergence less than 0.001.

After adopting said method, the present invention compared with prior art, have following remarkable advantage and beneficial effect: the present invention is based on the Kinect camera, adopted up-to-date human body attitude recognizer, also can reach on the commercial hardware more than per second 200 frames, under complex background and personage's attitude condition, can accurately follow the tracks of and cut apart face image, estimate that its head 3 d pose and degree of confidence thereof are in order to instruct the global shape conversion of Depth-AAM, the Depth-AAM algorithm is with texture image and exist the depth information of error to be combined into the four-way information of RGBD in order to train the Depth-AAM apparent model, and the four-way information of RGBD all inputted as the Depth-AAM algorithm data, use image pyramid algorithm and inverse compositional algorithm to accelerate iterative process, thereby the face characteristic that carries out accurate robust is demarcated, so that the human body attitude that the present invention can provide the Kinect camera is analyzed data and depth data is fused in the Depth-AAM algorithm, form the face feature extraction method based on 2.5 dimension images.

As improvement, point is 68, and manual the demarcation according to the facial image outline line is standard, be that take the right eye canthus the 1st point, other 67 some positions also are to determine along the facial image profile is unique, demarcate like this, operand is little, more is conducive to of the present inventionly efficiently carry out.

As improvement, described step is the facial image I under the Plays attitude 4. _i' resolution sizes is 42*43, apparent vectorial A _iResolution sizes also is 42*43, and like this, operand is little, more is conducive to of the present inventionly efficiently carry out.

As improvement, described step only has form parameter p and the global shape parameter q of the acquiescence that primary Depth-AAM match obtains in just using 5. in 6., for the first time later match all uses the form parameter p and the global shape parameter q that obtain after the last Depth-AAM match convergence to carry out initialization, like this, the faster face characteristic that carries out more accurately of the present invention is extracted.

As improvement, carry out the and 7. judged before the step, if head joint rotation direction θ is greater than 30 degree and direction degree of confidence Conf _θGreater than 0.8, then carried out for the 7. step, be about to be partitioned into the match of human body head RGBD four-way image substitution Depth-AAM algorithm iteration, otherwise just skipping the 7. goes on foot to finish face characteristic and extracts, like this, judged before iterative fitting whether target image is effective image, whether be recognizable image namely, for example, obtain if cut apart on the back side that the target image that obtains is behaved, then obviously this target image does not possess the people face part, can't be distinguished, above-mentionedly arrange that purpose is so that the present invention is effectively carried out, avoid invalid situation.

Description of drawings

Fig. 1 is the schematic flow sheet (Depth-AAM training algorithm) of the apparent model of employing principal component analysis method training Depth-AAM algorithm of the present invention.

Fig. 2 is that the apparent model of the Depth-AAM algorithm of finishing based on training of the present invention carries out the schematic flow sheet (Depth-AAM fitting algorithm) that face characteristic extracts.

Fig. 3 is 68 gauge point locations drawing of people's face of the present invention.

Fig. 4 is Depth-AAM triangulation network trrellis diagram of the present invention.

Embodiment

The invention will be further described below in conjunction with specific embodiment.

Fig. 3 is 68 gauge point locations drawing of people's face of the present invention.Face coincidence part mark mark is 60,61,62,63,64,65 counterclockwise, and the face center is number mark 66.

A kind of face feature extraction method that the present invention proposes is based on the Depth-AAM algorithm, the Depth-AAM algorithm belongs to the face characteristic location algorithm to two dimensional image---the improvement of AAM algorithm, the human body attitude that taking full advantage of the Kinect camera provides is analyzed data and depth data, they are fused in the AAM algorithm, form the human face characteristic positioning method based on 2.5 dimension images.

Described face feature extraction method may further comprise the steps,

1) adopt the principal component analysis method to train the apparent model of Depth-AAM algorithm:

1. utilize the training of Kinect collected by camera texture image and the depth image of facial image, depth image is compressed to 0 ~ 255 pixel coverage from 0 ~ 65535 pixel coverage, the α passage of substitution four-way image, merge into RGBD four-way image with texture image again, and it is carried out craft demarcate several point;

2. define people's face shape for forming v apex coordinate s=(x of grid ₁, y ₁..., x _v, y _v) ^TThe shape vector that the summit consists of is set up the two-dimensional linear model with the principal component analysis method, and shape vector is expressed as basic configuration s ₀Add m shape vector s _iLinear combination P=(p ₁..., p _m) ^TThe feature value vector of form matrix, s ₀Be the standard attitude of facial image, s _iEigenwert p _iThe characteristic of correspondence vector;

3. with s ₀, RGBD four-way image I _iWith its corresponding manual markings s _i ^*Be transformed into RGBD four-way facial image under the standard attitude, i.e. s with the piecemeal affined transformation ₀With s _i' triangle gridding corresponding one by one, as shown in Figure 4, piecemeal affined transformation expression formula is x'=a ₁X+a ₂Y+a ₃And y'=b ₁X+b ₂Y+b ₃, (x, y) is s ₀A upper coordinate, (x ', y ') be s _i' upper the coordinate corresponding with (x, y), a ₁With b ₂Be the convergent-divergent yardstick of directions X and Y-direction, a ₂And b ₁Be rotation yardstick, a ₃And b ₃Be the translation size of directions X and Y-direction, each triangle only need to be brought their three summits separately into the affined transformation expression formula, does not need each pixel is calculated, and adopts the method for undetermined coefficients to obtain corresponding parameter (a ₁, a ₂, a ₃, b ₁, b ₂, b ₃);

4. 3. all training facial images are gone on foot conversion through the, obtain its facial image I under the standard attitude _i', and adopt the principal component analysis method

Here λ _iThe parameter of i apparent vector, apparent parameter vector λ={ λ ₁, λ ₂..., λ _nThat input picture is corresponding to the eigenwert of the apparent parameter of this AAM model, to represent the full detail of input picture, i apparent vectorial A _i(x) corresponding to the large eigenwert of i in the apparent parameter vector;

Obtain A by 1. 2. 3. 4. going on foot training ₀, A ₁, A ₂Namely finish the training of the apparent model of Depth-AAM algorithm Deng the eigenwert of each appearance features;

The apparent model of the Depth-AAM algorithm of 2) finishing based on training carries out face characteristic and extracts:

5. the Kinect camera adopts the API of Kinect winsdk to be partitioned into human body image, and obtains head node location coordinate, head joint rotation direction θ and degree of confidence Conf thereof according to the human body depth image _θAccording to the human body depth image, human region is put white, zone outside the human body is put black, node location constantly enlarges the hunting zone about up and down simultaneously from the head, when all reaching black border about upper, stops the search of lower boundary, and the maximum region of definite people's face, its top left corner apex coordinate is designated as (x _HeadLU, y _HeadLU), zone length and width are designated as (1ength _Head, width _Head); In this specific embodiment, head node location and near depth information adopt the probe searching algorithm, expand to whole head zone, posting field position (t _x, t _y) and XY direction length, be used for initialization global shape function;

6. the global shape parameter q is defined as $N(x;q)=\left[\begin{array}{cc}1+a& -b\\ b& 1+a\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]+\left[\begin{array}{c}{t}_x\\ t_y\end{array}\right],$ Parameter (a, b) is expressed as a=kcos θ-and b=ksin θ, (t _x, t _y) be the translation of directions X and Y-direction, in order to write conveniently (a, b, t _x, t _y) be designated as (q ₁, q ₂, q ₃, q ₄) be the global shape parameter q; Obtain the target facial image during 5. purpose incites somebody to action exactly and carry out zooming and panning, compare with the standard attitude that 1. 2. 3. 4. goes on foot the apparent model that obtains;

7. the objective function of Depth-AAM algorithm match is the absolute value of input picture and formal synthesis image difference $E=\underset{x\∈s_0}{\mathrm{\Σ}}[A_0\left(x\right)+\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}_i{A}_i\left(x\right)-I\left(N(W(x;p);q)\right)],$ Rotation direction θ initialization q is used in primary Depth-AAM match ₁, q ₂, the people's face positional information initialization q that obtains in the use 5. ₃, q ₄P represents the standard attitude after the initialization, finding the solution parameter p and q makes the coordinate figure of the poor described point that hour obtains of image energy and apparent parameter vector namely finish face characteristic to extract, be specially, to parameter p and q differentiate, try to achieve variation delta p and the Δ q of parameter p and q, it is poor that iteration is asked for minimum image energy $\mathrm{\Δq}=H^{-1}\underset{i=0}{\overset{n}{\mathrm{\Σ}}}S{D}_i[I\left(N(W(x;p);q)\right)-A_0\left(x\right)],$ The scope of i is i=1 ..., 4, $\mathrm{\Δp}=H^{-1}\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{\mathrm{SD}}_{j+4}[I\left(N(W(x;p);q)\right)-A_0\left(x\right)],$ The scope of j is j=1. .., and, wherein

The scope of k is k=1 ..., 4,

The scope of l is l=1 ..., 68, H is the Hessian matrix, $H=\underset{x\∈S_0}{\mathrm{\Σ}}{\left[{\mathrm{SD}}_k\left(x\right)\right]}^T\left[{\mathrm{SD}}_k\left(x\right)\right].$

Point is 68, and manual the demarcation according to the facial image outline line is standard, is that take the right eye canthus the 1st point, other 67 some positions also are to determine that along the facial image profile is unique the calibration point position as shown in Figure 3.

Owing to directly use form parameter q at random not high to the precision that the facial image that is partitioned into carries out the Depth-AAM match, be difficult to reach fast convergence, so adopt the pyramid algorith of layering, will cut apart the facial image that obtains and narrow down to

With

Right first

The target facial image of size carries out primary Depth-AAM match, obtains rough form parameter p ¹With overall deformation parameter q, with primary Depth-AAM fitted shapes parameter p ¹Amplify 2 times, substitution Depth-AAM carries out the match second time and obtains form parameter p ²With overall deformation parameter q ², again with p ²Amplify 2 times, substitution Depth-AAM carries out for the third time match and obtains p ³, p ³68 point coordinates that namely match obtains, vectorial λ is exactly apparent parameter; Described pyramid algorith is divided into 3 layers, and every layer of pyramid maximum iteration time is 30 times, if the modulus of the difference of 2 p in front and back is considered as iteration convergence less than 0.001; After for the third time Depth-AAM match convergence, calculate apparent parameter and be

Described step is the facial image I under the Plays attitude 4. _i' resolution sizes is 42*43, apparent vectorial A _iResolution sizes also is 42*43.

Described step only has form parameter p and the global shape parameter q of the acquiescence that primary Depth-AAM match obtains in just using 5. in 6., and for the first time later match all uses the form parameter p and the global shape parameter q that obtain after the last Depth-AAM match convergence to carry out initialization.

Carry out the and 7. judged before the step, if head joint rotation direction θ is greater than 30 degree and direction degree of confidence Conf _θGreater than 0.8, then carried out for the 7. step, be about to be partitioned into the match of human body head RGBD four-way image substitution Depth-AAM algorithm iteration, 7. go on foot to finish face characteristic and extract otherwise just skip the.

Claims (6)

1. a face feature extraction method is characterized in that, may further comprise the steps,

1) adopt the principal component analysis method to train the apparent model of Depth-AAM algorithm:

1. utilize the training of Kinect collected by camera texture image and the depth image of facial image, depth image is compressed to 0 ~ 255 pixel coverage from 0 ~ 65535 pixel coverage, the α passage of substitution four-way image, merge into RGBD four-way image with texture image again, and it is carried out craft demarcate several point;

2. define people's face shape for forming v apex coordinate s=(x of grid ₁, y ₁..., x _v, y _v) ^TThe shape vector that the summit consists of is set up the two-dimensional linear model with the principal component analysis method, and shape vector is expressed as basic configuration s ₀Add m shape vector s _iLinear combination P=(p ₁..., p _m) ^TThe feature value vector of form matrix, s ₀Be the standard attitude of facial image, s _iEigenwert p _iThe characteristic of correspondence vector;

3. with s ₀, RGBD four-way image I _iWith its corresponding manual markings s _i ^*Be transformed into RGBD four-way facial image under the standard attitude, i.e. s with the piecemeal affined transformation _{0 with}s _i' triangle gridding corresponding one by one, piecemeal affined transformation expression formula is x'=a ₁X+a ₂Y+a ₃And y'=b ₁X+b ₂Y+b ₃, (x, y) is s ₀A upper coordinate, (x ', y ') be s _i' upper the coordinate corresponding with (x, y), a ₁With b ₂Be the convergent-divergent yardstick of directions X and Y-direction, a ₂And b ₁Be rotation yardstick, a ₃And b ₃Translation size for directions X and Y-direction adopts the method for undetermined coefficients to obtain corresponding parameter (a ₁, a ₂, a ₃, b ₁, b ₂, b ₃);

4. 3. all training facial images are gone on foot conversion through the, obtain its facial image I under the standard attitude _i', and adopt the principal component analysis method

Here λ _iThe parameter of i apparent vector, apparent parameter vector λ={ λ ₁, λ ₂..., λ _nThat input picture is corresponding to the eigenwert of the apparent parameter of this AAM model, to represent the full detail of input picture, i apparent vectorial A _i(x) corresponding to the large eigenwert of i in the apparent parameter vector;

By 1. 2. 3. 4. the step trains the eigenwert that obtains each appearance features namely to finish the training of the apparent model of Depth-AAM algorithm;

The apparent model of the Depth-AAM algorithm of 2) finishing based on training carries out face characteristic and extracts:

5. the Kinect camera adopts the API of Kinect winsdk to be partitioned into human body image, and obtains head node location coordinate, head joint rotation direction θ and degree of confidence Conf thereof according to the human body depth image _θAccording to the human body depth image, human region is put white, zone outside the human body is put black, node location constantly enlarges the hunting zone about up and down simultaneously from the head, when all reaching black border about upper, stops the search of lower boundary, and the maximum region of definite people's face, its top left corner apex coordinate is designated as (x _HeadLU, y _HeadLU), zone length and width are designated as (length _Head, width _Head);

6. the global shape parameter q is defined as $N(x;q)=\left[\begin{array}{cc}1+a& -b\\ b& 1+a\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]+\left[\begin{array}{c}{t}_x\\ t_y\end{array}\right],$ Parameter (a, b) is expressed as a=kcos θ-and b=ksin θ, (t _x, t _y) be the translation of directions X and Y-direction, in order to write conveniently (a, b, t _x, t _y) be designated as (q ₁, q ₂, q ₃, q ₄) be the global shape parameter q;

7. the objective function of Depth-AAM algorithm match is the absolute value of input picture and formal synthesis image difference

Rotation direction θ initialization q is used in primary Depth-AAM match ₁, q ₂, the people's face positional information initialization q that obtains in the use 5. ₃, q ₄, find the solution p and q and make the coordinate figure of the poor described point that hour obtains of image energy and apparent parameter vector namely finish face characteristic to extract, be specially, to parameter p and q differentiate, try to achieve variation delta p and the Δ q of parameter p and q, it is poor that iteration is asked for minimum image energy $\mathrm{\Δq}=H^{-1}\underset{i=0}{\overset{n}{\mathrm{\Σ}}}S{D}_i[I\left(N(W(x;p);q)\right)-A_0\left(x\right)],$ The scope of i is i=1 ..., 4, $\mathrm{\Δp}=H^{-1}\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{\mathrm{SD}}_{j+4}[I\left(N(W(x;p);q)\right)-A_0\left(x\right)],$ The scope of j is j=1. .., and, wherein

The scope of k is k=1 ..., 4,

The scope of l is l=1 ..., 68, H is the Hessian matrix, $H=\underset{x\∈S_0}{\mathrm{\Σ}}{\left[{\mathrm{SD}}_k\left(x\right)\right]}^T\left[{\mathrm{SD}}_k\left(x\right)\right].$

2. face feature extraction method according to claim 1 is characterized in that, point is 68, and manual the demarcation according to the facial image outline line is standard, is that take the right eye canthus the 1st point, other 67 some positions also are to determine along the facial image profile is unique.

3. face feature extraction method according to claim 1 is characterized in that, adopts the pyramid algorith of layering, will cut apart the facial image that obtains and narrow down to

With Right first

The target facial image of size carries out primary Depth-AAM match, obtains rough form parameter p ¹With overall deformation parameter q, with primary Depth-AAM fitted shapes parameter p ¹Amplify 2 times, substitution Depth-AAM carries out the match second time and obtains form parameter p ²With overall deformation parameter q ², again with p ²Amplify 2 times, substitution Depth-AAM carries out for the third time match and obtains p ³, p ³68 point coordinates that namely match obtains, vectorial λ is exactly apparent parameter; Described pyramid algorith is divided into 3 layers, and every layer of pyramid maximum iteration time is 30 times, if the modulus of the difference of 2 p in front and back is considered as iteration convergence less than 0.001.

4. face feature extraction method according to claim 1 is characterized in that, described step is the facial image I under the Plays attitude 4. _i' resolution sizes is 42*43, apparent vectorial A _iResolution sizes also is 42*43.

5. face feature extraction method according to claim 1, it is characterized in that, described step only has form parameter p and the global shape parameter q of the acquiescence that primary Depth-AAM match obtains in just using 5. in 6., and for the first time later match all uses the form parameter p and the global shape parameter q that obtain after the last Depth-AAM match convergence to carry out initialization.

6. face feature extraction method according to claim 1 is characterized in that, carries out the and 7. judges before the step, if head joint rotation direction θ is greater than 30 degree and direction degree of confidence Conf _θGreater than 0.8, then carried out for the 7. step, be about to be partitioned into the match of human body head RGBD four-way image substitution Depth-AAM algorithm iteration, 7. go on foot to finish face characteristic and extract otherwise just skip the.

Images